fix(langchain): handle empty llm generations

langfuse/langchain/CallbackHandler.py

@@ -1327,12 +1327,7 @@ def on_llm_end(
             self._log_debug_event(
                 "on_llm_end", run_id, parent_run_id, response=response, kwargs=kwargs
             )
-            response_generation = response.generations[-1][-1]
-            extracted_response = (
-                self._convert_message_to_dict(response_generation.message)
-                if isinstance(response_generation, ChatGeneration)
-                else _extract_raw_response(response_generation)
-            )
+            extracted_response = self._extract_llm_result_response(response)
 
             llm_usage = _parse_usage(response)
 
@@ -1501,6 +1496,20 @@ def _log_debug_event(
             f"Event: {event_name}, run_id: {run_id}, parent_run_id: {parent_run_id}"
         )
 
+    def _extract_llm_result_response(self, response: LLMResult) -> Any:
+        for generation in reversed(response.generations):
+            if not generation:
+                continue
+
+            response_generation = generation[-1]
+            return (
+                self._convert_message_to_dict(response_generation.message)
+                if isinstance(response_generation, ChatGeneration)
+                else _extract_raw_response(response_generation)
+            )
+
+        return None
+
 
 def _extract_raw_response(last_response: Any) -> Any:
     """Extract the response from the last response of the LLM call."""

tests/unit/test_langchain.py

@@ -132,6 +132,44 @@ def test_llm_callback_exports_generation_span(langfuse_memory_client, get_span):
     )
 
 
+def test_llm_callback_ends_generation_with_empty_generations(
+    langfuse_memory_client, get_span, json_attr
+):
+    handler = CallbackHandler()
+    run_id = uuid4()
+    response = LLMResult(
+        generations=[[]],
+        llm_output={
+            "token_usage": {
+                "prompt_tokens": 2,
+                "completion_tokens": 0,
+                "total_tokens": 2,
+            },
+            "model_name": "empty-model",
+        },
+    )
+
+    handler.on_llm_start(
+        {"name": "EmptyLLM"},
+        ["hello"],
+        run_id=run_id,
+        invocation_params={"model_name": "empty-model"},
+    )
+    handler.on_llm_end(response, run_id=run_id)
+
+    langfuse_memory_client.flush()
+    span = get_span("EmptyLLM")
+
+    assert (
+        span.attributes[LangfuseOtelSpanAttributes.OBSERVATION_MODEL] == "empty-model"
+    )
+    assert json_attr(span, LangfuseOtelSpanAttributes.OBSERVATION_USAGE_DETAILS) == {
+        "prompt_tokens": 2,
+        "completion_tokens": 0,
+        "total_tokens": 2,
+    }
+
+
 def test_lcel_chain_exports_intermediate_chain_spans(
     langfuse_memory_client, get_span, find_spans
 ):